Skip to content

Add float16 support to relu op #9267

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Mar 21, 2018
Merged

Add float16 support to relu op #9267

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d60180a
inital commit
kexinzhao Mar 20, 2018
018f3bd
small fix
kexinzhao Mar 20, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions paddle/fluid/operators/activation_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -613,3 +613,14 @@ REGISTER_OP(swish, ops::ActivationOp, ops::SwishOpMaker, swish_grad,
ops::grad_functor<double>>);

FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);

REGISTER_OP_CPU_KERNEL(relu,
ops::ActivationKernel<paddle::platform::CPUDeviceContext,
ops::ReluFunctor<float>>,
ops::ActivationKernel<paddle::platform::CPUDeviceContext,
ops::ReluFunctor<double>>);
REGISTER_OP_CPU_KERNEL(
relu_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
ops::ReluGradFunctor<float>>,
ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
ops::ReluGradFunctor<double>>);
14 changes: 14 additions & 0 deletions paddle/fluid/operators/activation_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@ limitations under the License. */

#define EIGEN_USE_GPU
#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/platform/float16.h"

namespace ops = paddle::operators;

Expand All @@ -31,3 +32,16 @@ namespace ops = paddle::operators;
ops::grad_functor<double>>);

FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);

REGISTER_OP_CUDA_KERNEL(
relu, ops::ActivationKernel<paddle::platform::CUDADeviceContext,
ops::ReluFunctor<float>>,
ops::ActivationKernel<paddle::platform::CUDADeviceContext,
ops::ReluFunctor<double>>,
ops::ActivationKernel<paddle::platform::CUDADeviceContext,
ops::ReluFunctor<paddle::platform::float16>>);
REGISTER_OP_CUDA_KERNEL(
relu_grad, ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
ops::ReluGradFunctor<float>>,
ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
ops::ReluGradFunctor<double>>);
1 change: 0 additions & 1 deletion paddle/fluid/operators/activation_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -772,7 +772,6 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
__macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor); \
__macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor); \
__macro(exp, ExpFunctor, ExpGradFunctor); \
__macro(relu, ReluFunctor, ReluGradFunctor); \
__macro(tanh, TanhFunctor, TanhGradFunctor); \
__macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \
__macro(sqrt, SqrtFunctor, SqrtGradFunctor); \
Expand Down
28 changes: 25 additions & 3 deletions python/paddle/fluid/tests/unittests/test_activation_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@

import unittest
import numpy as np
import paddle.fluid.core as core
from op_test import OpTest
from scipy.special import expit

Expand Down Expand Up @@ -212,18 +213,39 @@ def test_check_grad(self):
class TestRelu(OpTest):
def setUp(self):
self.op_type = "relu"
x = np.random.uniform(-1, 1, [11, 17]).astype("float32")
self.dtype = np.float32
self.init_dtype()

x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)
# The same reason with TestAbs
x[np.abs(x) < 0.005] = 0.02
self.inputs = {'X': x}
self.outputs = {'Out': np.maximum(self.inputs['X'], 0)}
out = np.maximum(x, 0)

self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
self.outputs = {'Out': out}

def test_check_output(self):
self.check_output()

def test_check_grad(self):
if self.dtype == np.float16:
return
self.check_grad(['X'], 'Out', max_relative_error=0.007)

def init_dtype(self):
pass


class TestFP16Relu(TestRelu):
def init_dtype(self):
self.dtype = np.float16

def test_check_output(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
Copy link
Contributor

@helinwang helinwang Mar 21, 2018 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just to double check: will the condition ever become True?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It will be true if the compute capability of GPU is >= 5.3, which means all pascal and volta GPUs will be tested.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, thank you!

self.check_output_with_place(place, atol=1e-3)


class TestBRelu(OpTest):
def setUp(self):
Expand Down